Open Source Agent Tools: October 16, 2026

1. CopilotKit Launches OpenMuse, a Self-Hostable Personal Agent

CopilotKit. CopilotKit launched OpenMuse on October 15, an open-source, self-hostable personal assistant that connects to any agent harness through the AG-UI protocol. The agent gets its own persistent Chromium browser, an optional Linux terminal, and a file workspace, and users can watch it work and take over the same browser session. It creates durable task plans with progress tracking, approvals, and receipts, integrates with Gmail and Calendar (with review required before sends and changes), and runs on iOS, Android, and the web. The project is MIT-licensed and currently in alpha. Source

2. Graphicle Brings Cited, Local Memory to Coding Agents

Graphicle. Graphicle, posted to Show HN this week, is a Rust port of the Graphify code-graph engine that adds persistent memory for coding agents. It indexes a repository (respecting .gitignore and skipping secrets), along with project docs, decisions, and captured agent sessions, then returns ranked excerpts with file and line citations when an agent calls its tools. Memories proposed by agents wait for user approval and go stale when the function they reference changes. It runs locally with no API key, account, or telemetry, connects to Claude Code, Codex, Cursor, VS Code, Claude Desktop, and any MCP client, and is licensed under Apache 2.0. Source

3. Endo Combines Existing Coding Agents in a “Higher-Order Harness”

Longsurf. Endo, shown on Hacker News on October 15, is a harness that treats existing agents such as Codex, Claude Code, and Antigravity as building blocks. Users can switch agents mid-thread with full conversation history passed along, write TypeScript workflow files that call agents from different providers in parallel and on a schedule, and branch side conversations from part of an answer. It uses existing subscriptions and CLI logins rather than API keys, stores sessions locally, ships as a macOS (Apple Silicon) app or from source, and is MIT-licensed at version 0.1.x. Source

4. BlastLoop Publishes a Structure-First Method for Directing Coding Agents

BlastLoop. BlastLoop, shown on Hacker News on October 15, is a process framework that splits AI coding work into three build altitudes (system, subsystem, and unit level), each with its own prompt role. Output then passes through structural, bug/ops, and adversarial test stages in fixed order, and structural problems found in the adversarial stage restart the loop and route back to the right altitude. The repository ships role prompt files, an AGENTS.md routing guide, and a handoff document, and is MIT-licensed at version 1.0.0. Source